Dave Hansen [hansendc@us.ibm.com] wrote:
| On Thu, 2007-05-31 at 15:48 +0400, Pavel Emelianov wrote:
| > > +void proc_flush_task(struct task_struct *task)
| > > +{
| > > +     int i;
| > > +     struct pid *pid;
| > > +     struct upid* upid;
| > > +
| > > +     pid = task_pid(task);
| > > +     if (!pid)
| > > +             return;
| > 
| > The code below will never be called as task flushes all his pids
| > in __unhash_process() that happens before this. Or did I miss smth? 
| 
| Nope, that's a very nice catch.  Suka is working on a patch to fix this
| right now.  We just need to keep the list of pids accessible a little
| longer, but I think Suka ran into a chicken-and-egg problem while trying
| to solve this.  Suka, care to post your workaround?

Yes. the chicken and egg problem is in the order of the __unhash_process()
and proc_flush_task().

With __unhash_process() first, proc_flush_task() has no way of knowing
the pid namespaces the process belongs to. With proc_flush_task() first,
we find the process but have its dentry removed from proc. This crashes
when I run a test in a tight loop.

A quick/dirty fix is to save the pid namespace list before __unhash_process()
and use that in proc_flush_task(). That would make the exit() path quite
expensive. I am still investigating the crash and looking for a better option.


| 
| -- Dave
_______________________________________________
Containers mailing list
Containers@lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers